only rebuild changed methods on PRs

[lacava]

depends on #220 - don't merge this until that lands, otherwise #220 stops being a clean promote of the docs/guardrail work.

every push and PR currently rebuilds all 27 images. the gate in build-and-test is hardcoded to should_run=true (5b13029), and check-changes computes what changed but nothing consumes it. the cost lands on contributors iterating: #209 went through 16 full 27-image runs, #210 and #212 another 24 between them.

pull requests now build only the methods they touch. everything else still rebuilds everything:

event	behaviour
pull_request	only changed methods
push to master/dev/docker-compose	full rebuild
weekly schedule (mon 06:00 UTC)	full rebuild
manual dispatch	full rebuild

the schedule is the part that matters most here. several methods install from git or unpinned pip and break with nobody touching the repo - that's calendar-driven, so during a quiet stretch there are no merges to catch it. that's roughly what happened in late july.

details:

• a method rebuilds if either algorithms/<name>/ or experiment/methods/<name>/ changed. the second matters: a regressor.py edit has to retest the method even though the install is untouched.
• changes to shared build inputs (dockerfiles, base_environment, scripts, entry.sh, configure.sh, workflows) still rebuild everything.
• build-and-test always runs and always reports for every algorithm, so the check names stay present and can be marked required. only the docker build step is skipped.
• dropped always() from build-and-test. with the gate inside the job, a failed check-changes would have left an empty build list, skipped every build, and reported green.
• check-changes no longer diffs against github.event.before, so a force-push to a CI branch no longer fails it. that's the red X that showed up on dev earlier today.

also fixes a long-standing bug: changed-experiments used awk field $2 on experiment/methods/<name>/..., which is the literal string methods, not the method name. it needs $3. nothing consumed that output before so it never surfaced.

the selection logic was checked against these cases before pushing:

new method, both dirs        -> smgp
regressor.py only            -> feat
install only                 -> operon
legacy dir with no image     -> (none)
tuned subdir                 -> (none)
docs only                    -> (none)
shared build input           -> FULL
two methods + a doc          -> pysr,tir
same method in both dirs     -> feat (no dupe)